Skip to content

multi-token prediction

Predict several future tokens in one forward pass

Visit official site
Category AI models

What multi-token-prediction is

The model is described as a Facebook research technique that generates several tokens in one forward pass.

What you can do with it

  • Train and run inference for non-commercial research
  • Encode input data with the supplied Tokenizer

Getting started

  • Install torch, fairscale, fire and sentencepiece
  • Unpack a checkpoint such as 7B_200B_4 and load the model

Pros & cons

✓ What we like

  • Extra prediction heads can skip standard autoregressive inference
  • A Tokenizer is supplied for model input

! What to watch out for

  • The stated licence is limited to non-commercial research
  • Local setup needs torch, fairscale, fire and sentencepiece

FAQ

How many tokens are predicted?

The material states multiple future tokens in a single forward pass.

What licence applies?

The material states free non-commercial research use and the LLaMA Acceptable Use Policy.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review